Highly accurate protein structure prediction for the human proteome

Kathryn Tunyasuvunakool, Jonas Adler, Zachary Wu, Tim Green, Michał Zieliński, Augustin Žídek, Alex Bridgland, Andrew Cowie, Clemens Meyer, Agata Laydon, Sameer Velankar, Gerard J. Kleywegt, Alex Bateman, K Taki, Alexander Pritzel, Michael Figurnov, Olaf Ronneberger, Russ Bates, Simon Köhl, Anna Potapenko, Andrew J. Ballard, Bernardino Romera‐Paredes, Stanislav Nikolov, Rishub Jain, Ellen Clancy, David Reiman, Stig Petersen, Andrew Senior, Koray Kavukcuoglu, Ewan Birney, Pushmeet Kohli, John Jumper, Demis HassabisNature 2021 · DOI

Abstract

Protein structures can provide invaluable information, both for reasoning about biological processes and for enabling interventions such as structure-based drug development or targeted mutagenesis. After decades of effort, 17% of the total residues in human protein sequences are covered by an experimentally determined structure. Here we markedly expand the structural coverage of the proteome by applying the state-of-the-art machine learning method, AlphaFold, at a scale that covers almost the entire human proteome (98.5% of human proteins). The resulting dataset covers 58% of residues with a confident prediction, of which a subset (36% of all residues) have very high confidence. We introduce several metrics developed by building on the AlphaFold model and use them to interpret the dataset, identifying strong multi-domain predictions as well as regions that are likely to be disordered. Finally, we provide some case studies to illustrate how high-quality predictions could be used to generate biological hypotheses. We are making our predictions freely available to the community and anticipate that routine large-scale and high-accuracy structure prediction will become an important tool that will allow new questions to be addressed from a structural perspective.

Introduction

Proteins are the fundamental workhorses of biology, with their 3D shapes dictating their biological functions. Historically, determining these shapes through experimental methods like X-ray crystallography or Cryo-EM has been a slow and resource-intensive process. Consequently, only about 17% of residues in the human proteome had been mapped to an experimental structure prior to this study. This "structural gap" has long hindered our understanding of disease mechanisms and the development of targeted therapeutics. Researchers from DeepMind and EMBL-EBI have addressed this challenge by scaling the AlphaFold2 deep learning system to provide high-accuracy structural predictions for 98.5% of human proteins.

Methodology and Validation

The team utilized AlphaFold2, a transformer-based machine learning model that integrates physical and biological knowledge of protein folding. A critical advancement in this study was the focus on Full-Chain Prediction, processing sequences up to 2,700 residues to predict how independent domains pack together.

To ensure reliability, the authors developed and validated two primary confidence metrics:

  • pLDDT (predicted Local Distance Difference Test): A per-residue score where values > 90 indicate accuracy comparable to high-resolution experimental structures, and values > 70 indicate a generally correct backbone.
  • pTM (predicted TM-score): A metric used to estimate the accuracy of global domain packing and orientation.
Validation of AlphaFold confidence metrics and proteome-wide coverage. Panel (a) shows the strong correlation between predicted (pLDDT) and actual (lDDT-Cα) accuracy. Panel (d) illustrates the high coverage achieved across various functional protein categories.

Key Results and Coverage

The application of AlphaFold2 to the human proteome has more than doubled the number of residues covered by high-confidence structures. Key findings include:

  • Massive Coverage: Predictions were generated for 98.5% of the human proteome, with 58% of all residues predicted with confidence (pLDDT > 70) and 36% with very high confidence (pLDDT > 90).
  • Structural Novelty: The dataset provides high-quality models for 1,290 proteins that previously had no structural template in the Protein Data Bank (PDB).
  • Global Accuracy: The model's ability to predict the overall topology of full chains was validated against experimental data, showing that the predicted TM-score is a reliable indicator of global structure quality.
Correlation between the predicted full-chain TM-score and the actual TM-score, demonstrating the model's ability to accurately estimate its own global prediction quality.

Key Insights: Confidence as a Biological Signal

A significant insight from this work is that low-confidence regions (pLDDT < 50) are often not model failures but rather indicators of intrinsically disordered regions—parts of proteins that lack a fixed 3D shape in isolation. Furthermore, AlphaFold demonstrated exceptional performance on membrane proteins, which are notoriously difficult to study experimentally. By providing full-chain predictions, the model allows researchers to generate new hypotheses regarding inter-domain interactions and complex signaling mechanisms.

Limitations and Future Directions

While this dataset represents a landmark achievement, several limitations remain. AlphaFold currently predicts single protein chains in isolation, whereas many proteins only achieve stability when part of a complex or bound to ligands. Additionally, the models provide static snapshots and do not yet capture the dynamic conformational changes essential for function. Future work will focus on predicting protein-protein complexes and further exploring the "dark proteome" of unstructured regions within their cellular contexts.

Summary updated

Abstract

Protein structures can provide invaluable information, both for reasoning about biological processes and for enabling interventions such as structure-based drug development or targeted mutagenesis. After decades of effort, 17% of the total residues in human protein sequences are covered by an experimentally determined structure<sup>1</sup>. Here we markedly expand the structural coverage of the proteome by applying the state-of-the-art machine learning method, AlphaFold<sup>2</sup>, at a scale that covers almost the entire human proteome (98.5% of human proteins). The resulting dataset covers 58% of residues with a confident prediction, of which a subset (36% of all residues) have very high confidence. We introduce several metrics developed by building on the AlphaFold model and use them to interpret the dataset, identifying strong multi-domain predictions as well as regions that are likely to be disordered. Finally, we provide some case studies to illustrate how high-quality predictions could be used to generate biological hypotheses. We are making our predictions freely available to the community and anticipate that routine large-scale and high-accuracy structure prediction will become an important tool that will allow new questions to be addressed from a structural perspective.
BibTeX
@inproceedings{tunyasuvunakool2021highly,
  author = {Kathryn Tunyasuvunakool and Jonas Adler and Zachary Wu and Tim Green and Michał Zieliński and Augustin Žídek and Alex Bridgland and Andrew Cowie and Clemens Meyer and Agata Laydon and Sameer Velankar and Gerard J. Kleywegt and Alex Bateman and K Taki and Alexander Pritzel and Michael Figurnov and Olaf Ronneberger and Russ Bates and Simon Köhl and Anna Potapenko and Andrew J. Ballard and Bernardino Romera‐Paredes and Stanislav Nikolov and Rishub Jain and Ellen Clancy and David Reiman and Stig Petersen and Andrew Senior and Koray Kavukcuoglu and Ewan Birney and Pushmeet Kohli and John Jumper and Demis Hassabis},
  title = {{Highly accurate protein structure prediction for the human proteome}},
  booktitle = {Nature},
  year = {2021},
  doi = {10.1038/s41586-021-03828-1},
}
Figure